Skip to content
Conference

Evaluation of Large-Scale Supercomputer Network Topology Using Betweenness Centrality

Sep 2026 · IEEE International Conference on Cluster Computing · pp. 254-264 · 0 citations · 9 references

Abstract

In large-scale supercomputers, efficient communication among compute nodes is crucial for the performance of parallel computing. To improve communication performance, many network topologies have been proposed and continue to be explored. Therefore, efficient topology exploration is essential during the design phase of post-exascale systems. To accelerate topology exploration, this paper presents an efficient evaluation method of network topologies based on graph theory. While prior graph-based approaches tend to focus on communication latency, often characterized by network diameter, we instead emphasize link-level bottlenecks caused by traffic congestion. This focus reflects the bandwidth-centric nature of modern large-scale parallel computing. We quantify communication bottlenecks at the link level using edge betweenness centrality. This metric is particularly suitable because it can be efficiently computed using the well-known Brandes' algorithm, enabling rapid topology exploration. Furthermore, we extend the method to support non-uniform communication patterns for more realistic evaluation by incorporating a traffic matrix into the computation. We show that the Brandes algorithm can be efficiently adapted for this purpose. Using the proposed method, we evaluate several representative topologies, including Full FatTree, Tapered FatTree, and Dragonfly, under both uniform and non-uniform traffic conditions. The results show that Full FatTree, despite its high network cost, achieves the smallest bottleneck. Dragonfly, which is known for its lower cost, demonstrates performance comparable to that of a 3:5 Tapered FatTree. We also demonstrate that our method enables designers to easily identify bottleneck links.

View source

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.